[ELF] Resolve defined symbols before undefined symbols

[MaskRay]

Summary

When parsing an object file, LLD interleaves undefined symbol resolution (which
may recursively fetch other lazy objects) with defined symbol resolution.

This may lead to surprising results, e.g. if an object file defines currently
undefined symbols and references another lazy symbol, we may interleave defined
symbols with the lazy fetch, potentially leading to the defined symbols
resolving to different files.

As an example, if both a.a(a.o) and a.a(b.o) define foo (not in COMDAT
group, or in different COMDAT groups) and __profd_foo (in COMDAT group
__profd_foo). LLD may resolve foo to a.a(a.o) and __profd_foo to
b.a(b.o), i.e. different files.

parse ArchiveFile a.a
  entry fetches a.a(a.o)
  parse ObjectFile a.o
    define entry
    define foo
    reference b
    b fetches a.a(b.o)
    parse ObjectFile b.o
      define prevailing __profd_foo
    define (ignored) non-prevailing __profd_foo

Assuming a set of interconnected symbols are defined all or none in several lazy
objects. Arguably making them resolve to the same file is preferable than making
them resolve to different files (some are lazy objects).

The main argument favoring the new behavior is the stability. The relative order
between a defined symbol and an undefined symbol does not change the symbol
resolution behavior. Only the relative order between two undefined symbols can
affect fetching behaviors.

---

The real world case is reduced from a Fuchsia PGO usage: a.a(a.o) has a
constructor within COMDAT group C5 while a.a(b.o) has a constructor within
COMDAT group C2. Because they use different group signatures, they are not
de-duplicated. It is not entirely whether Clang behavior is entirely conforming.

LLD selects the PGO counter section (__profd_*) from a.a(b.o) and the
constructor section from a.a(a.o). The __profd_* is a SHF_LINK_ORDER section
linking to its own non-prevailing constructor section, so LLD errors
sh_link points to discarded section. This patch fixes the error.

[MaskRay]

Add test

[jhenderson]

I've subscribed others from my team for visibility. My first thought was that changing symbol resolution order sounds a little scary. That's not to say this is undesirable though - I need to spend some time thinking about this. It might be helpful to have a graph drawn showing the different references that are relevant here and which definitions are in which groups, as I'm struggling to follow the words (I can probably piece it together, but it will require more effort).

[peter.smith]

The real world case is reduced from a Fuchsia PGO usage: a.a(a.o) has a constructor within COMDAT group C5
while a.a(b.o) has a constructor within COMDAT group C2. Because they use
different group signatures, they are not de-duplicated.

LLD selects the PGO counter section (profd_*) from a.a(b.o) and the
constructor section from a.a(a.o). The profd_* is a SHF_LINK_ORDER section
linking to its own non-prevailing constructor section, so LLD errors
sh_link points to discarded section. This patch fixes the error.

Can you write out the example in full, I'm not sure I got it. I would expect that if C5 and C2 have different signatures they should both be loaded as to the linker they are entirely separate entities. If the __profd_* sections have link order dependencies into group sections are not in the same group I think we have a problem with ELF file and not with the linker. Have I misunderstood the example?

Will need to have a think. I don't have any particular attachment to the way that we resolve undefined symbols today so I'm not going to object if there are other benefits.

[peter.smith]

OK. Had a chance to go through this in more detail. To make sure I understand; if the undefined symbols are in object files/libraries yet to be processed then there is no difference, but if there is a lazy symbol for an undefined then the object will be fetched immediately on seeing

// Handle global undefined symbols.
   if (eSym.st_shndx == SHN_UNDEF) {

Resulting in globals being loaded from a recursive chain of lazy loads before the globals from this object file have been loaded.

If I've got that right then I agree that this is a worthwhile change to make, if only because it makes library selection a bit more predictable/less-surprising.

[MaskRay]

In D95985#2541317, @peter.smith wrote:

The real world case is reduced from a Fuchsia PGO usage: a.a(a.o) has a constructor within COMDAT group C5
while a.a(b.o) has a constructor within COMDAT group C2. Because they use
different group signatures, they are not de-duplicated.

LLD selects the PGO counter section (profd_*) from a.a(b.o) and the
constructor section from a.a(a.o). The profd_* is a SHF_LINK_ORDER section
linking to its own non-prevailing constructor section, so LLD errors
sh_link points to discarded section. This patch fixes the error.

Can you write out the example in full, I'm not sure I got it. I would expect that if C5 and C2 have different signatures they should both be loaded as to the linker they are entirely separate entities. If the __profd_* sections have link order dependencies into group sections are not in the same group I think we have a problem with ELF file and not with the linker. Have I misunderstood the example?

Will need to have a think. I don't have any particular attachment to the way that we resolve undefined symbols today so I'm not going to object if there are other benefits.

parse ArchiveFile a.a
  entry fetches a.a(a.o)
  parse ObjectFile a.o
    define entry
    define foo
    reference b
    b fetches a.a(b.o)
    parse ObjectFile b.o
      define prevailing __profd_foo
    define (ignored) non-prevailing __profd_foo

Is the reduced example. The original case is (folks may not want to read it):

parse ArchiveFile libunwindstack.a
  _ZN11unwindstack8Unwinder6UnwindEPKNSt3__26vectorINS1_12basic_stringIcNS1_11char_traitsIcEENS1_9allocatorIcEEEENS6_IS8_EEEESC_ fetches libunwindstack.Unwinder.cpp.o
  parse ObjectFile libunwindstack.Unwinder.cpp.o
    _ZN11unwindstack3Elf12GetLastErrorEPNS_9ErrorDataE fetches libunwindstack.Elf.cpp.o
    parse ObjectFile libunwindstack.Elf.cpp.o
      _ZN11unwindstack12ElfInterface11GetLoadBiasI10Elf32_Ehdr10Elf32_PhdrEEmPNS_6MemoryE fetches libunwindstack.ElfInterface.cpp.o
        parse ObjectFile libunwindstack.ElfInterface.cpp.o
          _ZN11unwindstack19DwarfEhFrameWithHdrIjE12GetFdeFromPcEm fetches libunwindstack.DwarfEhFrameWithHdr.cpp.o
          parse ObjectFile libunwindstack.DwarfEhFrameWithHdr.cpp.o
            defines _ZN11unwindstack19DwarfEhFrameWithHdrIjEC2EPNS_6MemoryE
            references __profd__ZN11unwindstack19DwarfEhFrameWithHdrIjEC2EPNS_6MemoryE
              libunwindstack.ElfInterface.cpp.o defines __profd__ZN11unwindstack19DwarfEhFrameWithHdrIjEC2EPNS_6MemoryE

Let me abbreviate the long symbol to `xxx6MemoryE`.

If we let the ancestor parsing frame (libunwindstack.ElfInterface.cpp.o; simplified as b.o in my example) to define `__profd_xxx6MemoryE`, `xxx6MemoryE` and `__profd_xxx6MemoryE` will be defined by the same file.

xxx6MemoryE is not de-duplicated in libunwindstack.ElfInterface.cpp.o and libunwindstack.DwarfEhFrameWithHdr.cpp.o because the two object files use different group signature (C2 vs C5).

libunwindstack.DwarfEhFrameWithHdr.cpp.o
## C5 comdat
COMDAT group section [    3] `.group' [_ZN11unwindstack19DwarfEhFrameWithHdrIjEC5EPNS_6MemoryE] contains 2 sections:
   [Index]    Name
   [    4]   .text._ZN11unwindstack19DwarfEhFrameWithHdrIjEC2EPNS_6MemoryE
   [    5]   .rela.text._ZN11unwindstack19DwarfEhFrameWithHdrIjEC2EPNS_6MemoryE

libunwindstack.ElfInterface.cpp.o
## C2 comdat
COMDAT group section [    9] `.group' [_ZN11unwindstack19DwarfEhFrameWithHdrIjEC2EPNS_6MemoryE] contains 2 sections:
   [Index]    Name
   [   10]   .text._ZN11unwindstack19DwarfEhFrameWithHdrIjEC2EPNS_6MemoryE
   [   11]   .rela.text._ZN11unwindstack19DwarfEhFrameWithHdrIjEC2EPNS_6MemoryE

[MaskRay]

In D95985#2543028, @peter.smith wrote:

OK. Had a chance to go through this in more detail. To make sure I understand; if the undefined symbols are in object files/libraries yet to be processed then there is no difference, but if there is a lazy symbol for an undefined then the object will be fetched immediately on seeing

// Handle global undefined symbols.
   if (eSym.st_shndx == SHN_UNDEF) {

Resulting in globals being loaded from a recursive chain of lazy loads before the globals from this object file have been loaded.

If I've got that right then I agree that this is a worthwhile change to make, if only because it makes library selection a bit more predictable/less-surprising.

Yes. The idea is to postpone Undefined resolution (which may fetch lazy symbols) to the end. Previously, if we swap the order of a defined and an undefined in an object file, it can potentially change the symbol resolution result.
The new behavior is immune to defined/undefined shuffles.

Undefined shuffles (due to lazy symbols) can still result to different symbol resolution results.

[peter.smith]

I've put my approval down, as I think it is a worthwhile change to make regardless of whether the test case that had problems is strictly legal.

I think it will be worth waiting till Wednesday as that will have given people a week to object.

lld/ELF/InputFiles.cpp

1137

I've seen quite a few objects with more than 8 undefined symbols, particularly older projects. Is it worth running a quick experiment to see what the distribution is like? For example the undefined symbols per object in a clang/llvm build.

[phosek]

LGTM as well, I agree that this is valuable from the perspective of making the symbol resolution more predictable.

[MaskRay]

lld/ELF/InputFiles.cpp

1137

Linking clang:

• This function is called 2287 times.
• 228 elements have <= 8 elements.
• 465 elements have <= 16 elements.
• 842 elements have <= 32 elements.
• 1565 elements have <= 64 elements.

[jhenderson]

In general, I think this looks okay. I've spent some time investigating the behaviour of our propietary linker in comparison to this, and it looks look the modified LLD matches our linker's behaviour, so this seems like a good thing for improving compatibility.

That being said, a colleague and I have been looking at whether there are any potential problemas this behaviour change might cause. We should be able to conclude that investigation tomorrow. Could you hold on until end of UK workday on Thursday before pushing this, please?

lld/ELF/InputFiles.cpp

1204

[MaskRay]

Increase inline element number for unds
Improve comment

[jhenderson]

A colleague has just finished running a linker with this patch through our internal codebase and it seems to have been working fine. I haven't been able to come up with any realistic cases where this change will have made things worse, so LGTM.

lld/ELF/InputFiles.cpp

else if (sec == &InputSection::discarded)
           make<Undefined>(this, name, STB_LOCAL, eSym.st_other, type,
                           /*discardedSecIdx=*/secIdx);
     else
       this->symbols[i] = make<Defined>(this, name, STB_LOCAL, eSym.st_other,
                                        type, eSym.st_value, eSym.st_size, sec);
   }
 
   // Symbol resolution of non-local symbols.
+  SmallVector<size_t, 8> unds;
   for (size_t i = firstGlobal, end = eSyms.size(); i != end; ++i) {
     const Elf_Sym &eSym = eSyms[i];
     uint8_t binding = eSym.getBinding();
     if (binding == STB_LOCAL)
       continue; // Errored above.
 
     uint32_t secIdx = getSectionIndex(eSym);
     InputSectionBase *sec = this->sections[secIdx];
     uint8_t stOther = eSym.st_other;
     uint8_t type = eSym.getType();
     uint64_t value = eSym.st_value;
     uint64_t size = eSym.st_size;
     StringRefZ name = this->stringTable.data() + eSym.st_name;
 
     // Handle global undefined symbols.
     if (eSym.st_shndx == SHN_UNDEF) {
-      this->symbols[i]->resolve(Undefined{this, name, binding, stOther, type});
-      this->symbols[i]->referenced = true;
+      unds.push_back(i);
       continue;
     }
 
     // Handle global common symbols.
     if (eSym.st_shndx == SHN_COMMON) {
       if (value == 0 || value >= UINT32_MAX)
         fatal(toString(this) + ": common symbol '" + StringRef(name.data) +
               "' has invalid alignment: " + Twine(value));
if (binding == STB_GLOBAL || binding == STB_WEAK ||
         binding == STB_GNU_UNIQUE) {
       this->symbols[i]->resolve(
           Defined{this, name, binding, stOther, type, value, size, sec});
       continue;
     }
 
     fatal(toString(this) + ": unexpected binding: " + Twine((int)binding));
   }
+
+  // Undefined symbols (not those defined relative to non-prevailing sections)
+  // can trigger recursive fetch. Process defined symbols first so that a set of
+  // interconnected symbols will all be resolved to same file, instead of being
+  // resolved to different files.
+  for (size_t i : unds) {
+    const Elf_Sym &eSym = eSyms[i];
+    StringRefZ name = this->stringTable.data() + eSym.st_name;
+    this->symbols[i]->resolve(Undefined{this, name, eSym.getBinding(),
+                                        eSym.st_other, eSym.getType()});
+    this->symbols[i]->referenced = true;
+  }
 }
 
 ArchiveFile::ArchiveFile(std::unique_ptr<Archive> &&file)
     : InputFile(ArchiveKind, file->getMemoryBufferRef()),
       file(std::move(file)) {}
 
 void ArchiveFile::parse() {
   for (const Archive::Symbol &sym : file->symbols())

lld/test/ELF/interconnected-lazy.s

+# REQUIRES: x86
+
+# RUN: split-file %s %t
+# RUN: llvm-mc -filetype=obj -triple=x86_64 %t/main.s -o %t/main.o
+# RUN: llvm-mc -filetype=obj -triple=x86_64 %t/a.s -o %t/a.o
+# RUN: llvm-mc -filetype=obj -triple=x86_64 %t/b.s -o %t/b.o
+
+## foo and __foo are interconnected and defined in two lazy object files.
+## Test we resolve both to the same file.
+# RUN: ld.lld -y a -y foo -y __foo %t/main.o --start-lib %t/a.o %t/b.o --end-lib -o /dev/null | FileCheck %s
+
+# CHECK:      a.o: lazy definition of __foo
+# CHECK-NEXT: a.o: lazy definition of a
+# CHECK-NEXT: a.o: lazy definition of foo
+# CHECK-NEXT: b.o: definition of __foo
+# CHECK-NEXT: b.o: definition of foo
+# CHECK-NEXT: b.o: reference to a
+# CHECK-NEXT: a.o: definition of a
+
+#--- main.s
+.globl _start
+_start:
+  call b
+
+#--- a.s
+.globl a
+.weak foo
+a:
+foo:
+
+.weak __foo
+__foo:
+
+#--- b.s
+.globl b
+.weak foo
+b:
+  call a
+foo:
+
+.weak __foo
+__foo: